Research map / Image generation / Image restoration
Learning to Generate Realistic Noisy Images via Pixel-level Noise-aware Adversarial Training
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add pre-trained real denoiser D_d for image domain alignment instead of direct L1 on noisy images (ablation) | PNGAN · SIDD test, denoiser trained on generated data | PSNR: 14.54 -> 35.37 (RIDNet) | Helped |
| add pixel-level discriminator D (ablation) | PNGAN · SIDD test | PSNR: 35.37 -> 37.49 (RIDNet) | Helped |
| add perceptual loss L_p (ablation) | PNGAN · SIDD test | PSNR: 37.49 -> 37.92 (RIDNet) | Helped |
| use image-level discriminator instead of none (ablation) | PNGAN · SIDD test | PSNR | No clear effect |
| add multi-scale feature fusion to generator (ablation) | SMNet · SIDD test | PSNR: 35.62 -> 37.01 (RIDNet) | Helped |
| add Shift-Invariant Downsample to generator (ablation) | SMNet · SIDD test | PSNR: 37.01 -> 37.23 (RIDNet) | Helped |
| add Fast Channel Attention to generator (ablation) | SMNet · SIDD test | PSNR: 37.23 -> 37.92 (RIDNet) | Helped |
| apply PNGAN to synthetic setting1 (AWGN) | RIDNet trained from scratch · SIDD test | PSNR: 22.55 -> 37.92 | Helped |
| apply PNGAN to synthetic setting2 (CycleISP) | RIDNet trained from scratch · SIDD test | PSNR: 36.13 -> 38.71 | Helped |
| finetune denoiser with PNGAN-generated noisy pairs | MPRNet · SIDD | PSNR: 39.71 -> 40.06 | Helped |
| finetune denoiser with PNGAN-generated noisy pairs | MIRNet · PolyU | PSNR: 39.18 -> 40.55 | Helped |
| set finetuning ratio q of generated data to 0 (only SIDD real data) (ablation) | RIDNet · PolyU and Nam | PSNR | Mixed |
Closest papers by meaning
- Noise Robust Generative Adversarial Networks (CVPR 2020 · similarity 0.92)
- GAN2GAN: Generative Noise Learning for Blind Denoising with Single Noisy Images (ICLR 2021 · similarity 0.91)
- sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows (ICLR 2024 · similarity 0.91)
- End-to-End Unpaired Image Denoising with Conditional Adversarial Networks (AAAI 2020 · similarity 0.90)
- Image Blind Denoising With Generative Adversarial Network Based Noise Modeling (CVPR 2018 · similarity 0.90)
- Robust Image Denoising through Adversarial Frequency Mixup (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.